A method, device, and storage medium for simulating high gradient stress fields in porous structures based on hole-guided physical constraint learning.
By arranging high-density physical collocation points in the high-gradient neighborhood of the hole, using a radial weighting function and adaptive loss at partition boundaries, and combining finite element weak supervision, a hole-guided physical constraint learning model is established. This solves the problems of inaccurate prediction and high cost of repeated simulation in the high-gradient region of the hole, and achieves fast and reliable stress field simulation.
Patent Information
- Application Number
- CN202611236836.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-14
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies suffer from several problems, including insufficient physical points in the high gradient region around the hole, imbalance of constraints under different boundary conditions, the tendency of pure data-driven models to smooth out peak values around the hole, and high costs associated with repetitive finite element simulations. These issues make it difficult to meet the needs for rapid evaluation and frequent updates.
By arranging high-density physical collocation points in the high-gradient neighborhood of the hole edge, using the radial weighting function to weight the residual of the balance equation, constructing the partition boundary adaptive loss, and introducing weak supervision of the finite element hole edge, a hole-guided physical constraint learning model is established to improve the prediction reliability of the stress concentration region at the hole edge.
It improves the predictive reliability of peak stress and circumferential stress distribution at the hole edge, reduces the computational cost of repetitive finite element simulation under multiple working conditions, realizes millisecond-level stress field inference, and supports rapid evaluation and structural optimization under multiple working conditions.
Smart Images

Figure CN123069751A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of structural simulation, intelligent computing, and finite element-assisted modeling technology, specifically a method, device, and storage medium for simulating high gradient stress fields of porous structures based on hole-guided physical constraint learning. Background Technology
[0002] Hole structures are widely used in engineering equipment such as wind turbine generators, aerospace equipment, mechanical connection structures, flange connection structures, bearing connection structures, pressure vessels, and gearbox housings. In wind turbine generators, critical components such as pitch bearings, yaw bearings, main shaft flanges, and blade root connections are typically connected using a large number of high-strength bolts. The bolt holes alter the local geometric continuity and load transfer path.
[0003] Due to geometric discontinuities and load redistribution at the hole edges, localized high-gradient stress fields easily form in the ligamentous regions between holes and adjacent holes. The peak stress, peak angle position, and circumferential stress distribution at the hole edges are closely related to fatigue damage accumulation, crack initiation, fretting wear, and connection loosening. When the peak stress at the hole edges is underestimated or its position shifts, it may lead to misjudgments of structural safety margins and fatigue risks.
[0004] Finite element analysis (FEM) can account for finite dimensions, complex boundaries, material parameters, and local mesh refinement, making it an important tool for obtaining the full-field stress and stress distribution at the hole edges of porous structures. However, in scenarios involving multiple loads, multiple operating conditions, structural parameter updates, and digital twin state assessments, it is necessary to repeatedly perform high-fidelity FEM solutions on the same structure. To ensure the accuracy of the stress at the hole edges, the FEM model typically requires local mesh refinement at the hole edges, resulting in high modeling and computational costs for repeated simulations, making it difficult to meet the needs of rapid assessment and frequent updates.
[0005] Neural network surrogate models offer fast inference speeds, but purely data-driven models rely on a large number of high-fidelity samples. Since the peak value at the hole edge is located in a narrow, high-gradient region, insufficient hole edge samples can lead to smoothed predictions of peak stress, potentially resulting in distortion of the circumferential stress curve or non-physical oscillations. Improving accuracy by significantly increasing the number of finite element samples at the hole edge would also substantially increase data generation costs.
[0006] Physically constrained neural networks (PNNs) enhance the physical consistency of predictions under limited labeled data by incorporating governing equations and boundary conditions into the loss function. However, conventional PNNs typically employ uniformly distributed physical points across the entire computational domain and use fixed loss weights. For structures with holes, the response changes smoothly over a large area of the far field, while the high-gradient region around the hole is relatively small. Uniform point distribution can lead to the training process being dominated by the smooth far field region, and fixed boundary weights may also cause an imbalance between the outer boundary constraints and the free boundary constraints around the hole, thus affecting the prediction accuracy of the peak stress and peak position at the hole edge.
[0007] Therefore, it is necessary to provide a fast simulation method for high gradient regions at the edges of porous structures. This method improves the prediction reliability of stress concentration regions at the edges of holes by using hole-edge guided sampling, radial weighted physical constraints, adaptive weighting of partition boundaries, and weak supervision using a small amount of finite element hole-edge data. It also enables the trained model to be quickly extended to porous structures through parameter transfer and incremental training. Summary of the Invention
[0008] The purpose of this invention is to overcome the problems in the prior art, such as insufficient physical point allocation in the high gradient region of the hole edge, imbalance of constraints under different boundary conditions, easy smoothing of hole edge peaks in pure data-driven models, and high cost of repeated finite element simulations. The invention provides a method, device, and storage medium for simulating the high gradient stress field of porous structures based on hole-guided physical constraint learning.
[0009] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:
[0010] Step 1: Obtain the geometric parameters, material parameters, boundary load parameters, and finite element reference data for the hole edge of the structure to be simulated.
[0011] Step 2: Determine the computational domain, outer boundary, and hole boundary of the hole structure based on the geometric parameters, and expand the hole boundary outward with a preset radial width to construct a high gradient neighborhood around the hole edge.
[0012] Step 3: Arrange high-density physical points around the hole in the high-gradient neighborhood of the hole edge, and arrange background physical points in the background region outside the high-gradient neighborhood of the hole edge.
[0013] Step four: Establish a hole-guided physical constraint learning model with spatial coordinates as input and displacement components as output. Calculate strain components through automatic differentiation and stress components based on material constitutive relations.
[0014] Step 5: Substitute the predicted stress components into the static equilibrium equation to obtain the equilibrium equation residuals; construct a radial weighting function based on the radial distance from the physical collocation point to the hole boundary, and weight the equilibrium equation residuals.
[0015] Step 6: Divide the boundary of the porous structure into the outer boundary and the hole boundary, construct the outer boundary loss and the hole boundary loss respectively, and adaptively adjust the boundary loss weight according to the relative magnitude of the two types of boundary residuals.
[0016] Step 7: Extract finite element reference stress data at several discrete angular positions at the hole boundary, and construct a weak monitoring loss for hole edge stress based on the difference between the model-predicted stress and the finite element reference stress.
[0017] Step 8: Construct the total loss function based on the radial weighted balance equation loss, partition boundary loss, and weak supervision loss of hole edge stress, and use an optimization algorithm to train the hole-guided physical constraint learning model.
[0018] Step nine involves inputting the spatial coordinates of the area to be predicted into the trained hole-guided physical constraint learning model to obtain the displacement field, strain field, and stress field, and extracting the stress distribution at the hole edge, peak stress, peak position, and overall stress results.
[0019] Step 10: For porous structures, transfer the model parameters trained on single-pore structures to porous structures, rearrange high-density physical points in the edge regions of each pore and the ligament regions between pores, and perform incremental training based on the boundary conditions of the porous structure and the finite element reference data.
[0020] Furthermore, the hole-containing structure to be simulated is a single-hole plate, a double-hole plate, a multi-hole plate, a bolted connection structure, a flange connection structure, a bearing connection structure, a wind turbine pitch bearing bolt hole area, or other engineering structures with stress concentration problems at the hole edge.
[0021] Furthermore, the geometric parameters include at least one of the following: hole radius, hole center position, number of holes, distance between holes, structural length, structural width, and structural thickness; the material parameters include elastic modulus and Poisson's ratio; and the boundary load parameters include at least one of the following: displacement boundary, far-field tensile load, equivalent surface force, and free boundary condition.
[0022] Furthermore, for a hole radius of... A single-hole structure containing holes, wherein the high gradient neighborhood of the hole edge Represented as:
[0023]
[0024] In equation (1), Point radial distance to the center of the hole This represents the radial expansion width of the high gradient neighborhood around the hole. The value can be determined based on the hole radius, structural dimensions, finite element stress gradient variation characteristics, or engineering experience. In a preferred embodiment, The outer radius of the high gradient neighborhood around the hole is .
[0025] Furthermore, the two-dimensional displacement field output by the hole-guided physical constraint learning model is represented as:
[0026]
[0027] In equation (2), The two-dimensional displacement vector is composed of the two displacement components, and the two-dimensional displacement components output by the hole-guided physical constraint learning model are: and , To compute spatial coordinates within the domain along Displacement components in the direction, Spatial coordinates of this location along Displacement components in the direction, indicated by the upper right subscript. This indicates transpose.
[0028] Under small deformation conditions, the strain components are obtained through automatic differentiation:
[0029]
[0030] Under plane stress conditions, the stress components are obtained according to the linear elastic constitutive relation:
[0031]
[0032]
[0033]
[0034] In equations (3) to (6), For elastic modulus, Poisson's ratio; and They are respectively direction, Normal strain components in the direction, for Planar engineering shear strain components; and They are respectively direction, Normal stress components in the direction, for Shear stress components in a plane.
[0035] Furthermore, neglecting body forces, the two-dimensional static equilibrium equations are:
[0036]
[0037] Substituting the model's predicted stress into equation (7-1), we obtain... Directional equilibrium equation residuals and Directional equilibrium equation residuals The residual components are defined as follows:
[0038]
[0039] In equation (7-2), when the two-dimensional static equilibrium equations are completely satisfied, and All values are zero. A radial weighting function is constructed based on the radial distance from the physical collocation point to the hole boundary:
[0040]
[0041] In equation (8), The hole edge constraint reinforcement coefficient is... These are the radial attenuation control parameters; Let be the radial distance from any physical point to the center of the hole. Where is the radius of the hole. Let be the radial distance from the physical point to the boundary of the circular hole. This represents the radial weight corresponding to the physical alignment point. The radial weight increases as the physical alignment point approaches the hole boundary; it gradually decreases as the physical alignment point moves away from the hole boundary.
[0042] Furthermore, the radially weighted balance equation loss is:
[0043]
[0044] In equation (9), The total number of physical points. Indicates the first Spatial coordinates of a physical point This represents the radial distance from the physical point to the center of the hole. This represents the radial weight corresponding to the physical collocation point. and They represent the first At each physical coordinate point Directional equilibrium equation residuals and sum Residuals of directional equilibrium equations.
[0045] Furthermore, the outer boundary loss and the hole boundary loss are expressed as follows: and The partition boundary loss is:
[0046]
[0047] In the During the training round, the outer boundary weights and hole boundary weights are determined according to the first round. Mean boundary residuals after training round and
[0048] Update:
[0049]
[0050]
[0051] In equations (10)-(12), and These are the outer boundary loss and the hole boundary loss, respectively. For partition boundary loss; and The first The training round uses outer boundary loss weights and hole boundary loss weights; and The first The average value of the boundary residual amplitude at the outer boundary sampling point and the hole boundary sampling point after the round of training.
[0052] When a certain type of boundary residual is large, its corresponding weight will automatically increase in the next round of training, so that the optimization process pays more attention to the boundary conditions that have not yet been fully satisfied.
[0053] Furthermore, at the hole boundary Selecting Discrete angular positions And extract the finite element reference stress. ,in, Where is the radius of the hole. For the first The polar angle of a finite element reference point at the edge of a hole relative to the center of the hole. Indicates the type of target stress component being monitored; the weak monitoring loss for hole edge stress is:
[0054]
[0055] In equation (13), This indicates the position of the hole-guided physical constraint learning model at the hole edge. The predicted first Stress-like components, This represents the corresponding reference stress component obtained from the finite element model at the same hole edge location; the first... The stress components are selected from normal stress, circumferential stress, shear stress, or equivalent stress. The number of finite element reference points is much smaller than the number of physical collocation points and boundary collocation points, which can provide target shape constraints for the circumferential stress curve at the hole edge with less data cost, suppressing non-physical oscillations and smoothing peaks.
[0056] Furthermore, the total loss function of the hole-guided physical constraint learning model is:
[0057]
[0058] In equation (14), Let the total loss function be the physical constraint learning model guided by the hole. , and These are the weighting coefficients for the radial weighted equilibrium equation loss, the partition boundary loss, and the weak supervision loss of the hole edge stress, respectively.
[0059] Furthermore, the hole-guided physical constraint learning model employs a fully connected neural network containing a sinusoidal activation function to enhance its ability to express high-frequency and high-gradient responses at the hole edges; the Adam optimization algorithm is used to update the model parameters, and physical collocation points and boundary collocation points are randomly resampled in each training round.
[0060] Furthermore, the present invention also provides a simulation device for high gradient stress field of porous structures based on hole-guided physical constraint learning. The device includes a parameter acquisition module, a neighborhood construction module, a collocation generation module, a model construction module, a physical constraint construction module, a boundary constraint construction module, a finite element weak supervision module, a model training module, and a stress field output module.
[0061] Furthermore, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores computer instructions, and the processor executes the computer instructions to implement the above-described method; the present invention also provides a computer-readable storage medium storing the computer instructions. The present invention also provides a computer program product, the computer program product including a computer program or computer instructions, which, when executed by a processor, implement the above-described method.
[0062] The beneficial effects of this invention are:
[0063] (1) By increasing the density of physical points in the high gradient neighborhood of the hole edge, the limited training resources are concentrated in the stress concentration area of the hole edge, avoiding insufficient constraints in the key area of the hole edge due to uniform point allocation across the entire domain.
[0064] (2) By increasing the weight of the residual of the equilibrium equation near the hole edge through the radial weight function, the physical constraint strength of the peak stress region near the hole edge is enhanced, and the risk of peak stress being smoothed is reduced.
[0065] (3) By constructing losses for the outer boundary and the hole boundary respectively and performing adaptive weighting, the constraint imbalance between the far-field load boundary and the hole edge free boundary can be alleviated, and the training convergence stability can be improved.
[0066] (4) Weak supervision can be achieved by using a small number of finite element reference stresses at the hole edge, which can constrain the circumferential stress distribution and peak position at the hole edge without significantly increasing the number of finite element samples in the whole field, and suppress local non-physical oscillations.
[0067] (5) By transferring the stress concentration law at the hole edge learned in the single-hole structure to the porous structure, and performing incremental training in the hole edge and inter-hole ligament region, the computational cost required for de novo training of the porous structure is reduced.
[0068] (6) The trained model can complete stress field inference in milliseconds and can be used for rapid evaluation of multiple working conditions, structural optimization, fatigue risk analysis and digital twin model update. Attached Figure Description
[0069] Figure 1 This is a schematic diagram of the simulation method for high gradient stress field of porous structures based on hole-guided physical constraint learning;
[0070] Figure 2 A schematic diagram of the computational domain, hole boundary, and boundary load of a structure with holes;
[0071] Figure 3 A schematic diagram of the high gradient neighborhood and physical collocation point distribution around the hole;
[0072] Figure 4 A schematic diagram of the physical constraint learning model structure for hole guidance;
[0073] Figure 5 This is a schematic diagram of the radial weighted equilibrium equation constraints and partition boundary constraints;
[0074] Figure 6 Schematic diagram of weak supervision constraint construction for finite element method at the hole edge;
[0075] Figure 7 This is a schematic diagram of the training process for migrating from a single-pore structure to a porous structure.
[0076] Figure 8 The structural block diagram of a simulation device for high gradient stress field of porous structures based on hole-guided physical constraint learning;
[0077] Figure 9 This is a schematic diagram of the hardware structure of a computer device. Detailed Implementation
[0078] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description. The geometric parameters, material parameters, network parameters, and training parameters used in the following embodiments are only used to illustrate the implementation process of the present invention, and those skilled in the art can adjust them according to the structure to be simulated and the accuracy requirements.
[0079] like Figure 1 As shown, a simulation method for high gradient stress field of porous structures based on hole-guided physical constraint learning includes the following steps:
[0080] Step 1: Obtain the basic parameters of the porous structure to be simulated
[0081] Obtain the geometric parameters, material parameters, boundary load parameters, and finite element reference data for the hole-containing structure to be simulated. Geometric parameters are used to determine the computational domain and hole edge location; material parameters are used to establish the constitutive relationship between strain and stress; boundary load parameters are used to construct the external boundary constraints; and finite element reference data for the hole edge is used to construct the weakly monitored stress loss at the hole edge.
[0082] The object to be simulated can be a single-hole plate, a double-hole plate, a multi-hole plate, a local structure of bolt holes, a local structure of flange holes, or a local area around the hole in the bolt connection of a wind turbine pitch bearing. For circular hole structures, the geometric parameters include at least the hole radius R and the coordinates of the hole center; for multi-hole structures, the center positions of each hole and the distance between holes are also included.
[0083] Step 2: Establish the computational domain and boundary conditions for the porous structure.
[0084] like Figure 2 As shown, the computational domain Ω is established based on the outer boundary dimensions of the structure, and the hole boundary is determined based on the hole center position and hole radius. And denote the outer boundary of the structure as The outer boundary can be subjected to displacement boundary, far-field tensile load or equivalent surface force according to the actual problem; the hole boundary is set as a free boundary when there is no contact load, and the product of stress tensor and hole edge normal vector is zero.
[0085]
[0086] In equation (15), For stress tensor, Let the unit vector be the outward normal vector of the hole boundary. It is a zero vector; for local engineering models, the displacement and traction obtained from the previous finite element model, experimental measurement, or equivalent load calculation can also be used as the outer boundary input.
[0087] Step 3: Establish a finite element reference model and extract reference data from the hole edge.
[0088] A reference model corresponding to the porous structure to be simulated was established using finite element method (FEM) software. Local mesh refinement was applied to the areas around the holes and the ligaments between the holes, while a relatively coarse mesh was used in the background areas far from the holes. The FEM reference model was used to generate full-field verification data, and a small amount of reference stress was extracted at discrete angle positions around the holes.
[0089] The reference data at the hole edge can be the normal stress, circumferential stress, shear stress, or equivalent stress in the main loading direction. Only a small number of hole edge reference points are used during model training; the remaining finite element full-field results are used for accuracy verification after training, thus avoiding the degeneration of this invention into a purely data-driven method relying on a large number of full-field labels.
[0090] Step 4: Construct high-gradient neighborhoods and differentiated physical collocations around the hole edges.
[0091] like Figure 3 As shown, a high-gradient neighborhood of the annular hole edge is constructed outside the hole boundary. For the single-hole circular plate problem, the inner radius of the annular region is... The outer radius is High-density physical points are arranged in the high-gradient neighborhood of the hole according to radial distance and circumferential angle. Background physical points are generated in the background area using random sampling, grid sampling, Latin hypercube sampling or low-difference sequence sampling.
[0092] High-density physical collocation points at the pore edges are used to strengthen the equilibrium equation constraints in the high-gradient regions at the pore edges, while background physical collocation points are used to maintain global equilibrium consistency. For porous structures, in addition to the annular regions outside each pore, the collocation point density is also increased in the interpore ligament regions between adjacent pores.
[0093] Step 5: Establish a hole-guided physical constraint learning model
[0094] like Figure 4 As shown, the model uses spatial coordinates As input, in two-dimensional displacement components and The output is a fully connected neural network with a sinusoidal activation function. This improves the ability to represent high-frequency responses and steep gradients at the hole edge. The strain components are obtained by automatic differentiation, and the stress components are calculated according to equations (4) to (6).
[0095] The model outputs displacements instead of directly outputting stresses, ensuring consistency between the predicted displacements, strains, and stresses in terms of geometric and material constitutive relationships.
[0096] Step 6: Construct radially weighted equilibrium equation constraints
[0097] The residuals of the equilibrium equation at each physical collocation point are calculated according to equation (7). To avoid large-scale smooth background regions dominating training, the radial weight function shown in equation (8) is constructed based on the radial distance from the physical collocation point to the hole boundary, and the radial weighted equilibrium equation loss is constructed using equation (9).
[0098] Physical collocation points near the hole boundary have a larger weight, which makes the model prioritize reducing the residuals of the equilibrium equations in the stress concentration region near the hole. As the distance from the hole boundary increases, the weight gradually decreases, but background collocation points are still retained to ensure the physical consistency of the stress field across the entire domain.
[0099] Step 7: Construct partition boundary constraints
[0100] like Figure 5 As shown, the outer boundary and hole boundary They are processed separately. The outer boundary loss is used to constrain the displacement boundary or traction boundary, and the hole boundary loss is used to constrain the free hole boundary. The partition boundary loss is constructed according to Equation (10), and the weights of the two types of boundary losses are updated during training according to Equations (11) and (12).
[0101] When the residual of the free boundary at the hole edge is large, the weight of the hole boundary automatically increases; when the residual of the load on the outer boundary is large, the weight of the outer boundary automatically increases. This method can prevent the fixed boundary weights from becoming unbalanced at different training stages.
[0102] Step 8: Construct weak supervision constraints for finite element hole edges
[0103] like Figure 6 As shown, finite element reference points are selected at preset angle intervals on the hole boundary, and the finite element reference stress at the corresponding position is extracted. The weak monitoring loss of hole edge stress is constructed using equation (13). The finite element reference points can be evenly arranged, or the sampling density can be increased in the estimated peak sensitive area.
[0104] Weakly supervised data is used to provide high-fidelity anchor points for the hole edge stress curve, constraining the target shape and peak position near the peak. Since the number of reference points is much smaller than the total number of nodes in the field, this constraint does not significantly increase the cost of data generation and training.
[0105] Step 9: Construct the total loss function and train the model
[0106] The total loss function is constructed according to equation (14). The model parameters are randomly initialized at the start of training, and the Adam optimization algorithm is used for iterative updates. To avoid overfitting the model to the fixed set of collocations, the internal physical collocations and boundary collocations are randomly resampled in each round of training.
[0107] The training process allows setting a maximum number of training epochs, a loss convergence threshold, or a threshold for the rate of change of loss over a consecutive number of epochs as a stopping condition. After training, the model parameters are saved for quick lookup under different load conditions on the same structure.
[0108] Step 10: Output the simulation results of the high gradient stress field
[0109] By inputting the spatial coordinates of the region to be predicted into the trained model, the displacement field, strain field, and stress field are obtained. Based on the predicted stress components, the equivalent stress and circumferential stress at the hole edge can be calculated, and the peak stress and peak angle position can be extracted.
[0110]
[0111] As shown in Equation (16), the output results include at least one of the following: full-field stress cloud diagram, stress distribution at the hole edge, circumferential stress curve at the hole edge, peak stress, peak position, and stress concentration factor.
[0112] The stress concentration factor can be determined by the ratio of the peak stress at the hole edge to the corresponding nominal stress; the peak position can be represented by the polar angle of the peak point relative to the hole center.
[0113] Single-hole porous structure embodiment
[0114] The following uses a square plate with a central circular hole subjected to uniaxial tension as an example to illustrate the specific implementation process of the present invention, as shown in the following table of single-hole structure parameters: Square plate side length 100 mm thick The radius of the central hole is 10 mm. The thickness is 18 mm, and the material's elastic modulus is... 210 GPa, Poisson's ratio The far-field tensile stress is 0.30. The strength is 10 MPa; the finite element reference model uses a local mesh size of 0.5 mm at the hole edge and gradually coarsens the mesh in the far field region.
[0115] Single-hole structure parameter description table
[0116]
[0117] In this embodiment, the high gradient neighborhood around the hole edge Extend to The model employs a SIREN-type sinusoidal activated fully connected network and is trained for 1500 epochs using the Adam optimization algorithm. Internal physical collocation points and boundary collocation points are resampled in each training epoch. The finite element normal stress at discrete angle positions along the hole edge serves as weak supervision information; the full-field finite element results are used only for validation.
[0118] After training, 360 angular positions were uniformly selected at the hole edge to evaluate the normal stress in the principal loading direction. Compared with the finite element reference results, the root mean square error (RMSE) of the hole-guided physical constraint learning model in this embodiment is 0.3831 MPa, the mean absolute error (MAO) is 0.2867 MPa, the coefficient of determination (R²) is 0.9970, the peak absolute error is 0.1599 MPa, and the peak angle absolute error is 0.793°. These results demonstrate that the present invention can simultaneously improve the prediction reliability of the hole edge stress curve, peak magnitude, and peak position.
[0119] At the same evaluation point near the hole, a single finite element solution and result extraction takes approximately 36.0 seconds, while the trained model takes approximately 0.01157 seconds for a single inference iteration, corresponding to an inference speedup of approximately 3111 times. The training cost is a one-time cost, and the trained model can be used for repeated rapid evaluations of the same structure under different load conditions.
[0120] Porous Structure Migration Training Examples
[0121] like Figure 7 As shown, after the single-hole structure model is trained, the model parameters are used as the initial parameters for the double-hole or multi-hole structure model. In the double-hole structure, the two circular holes are symmetrically arranged along a set direction. The hole radius, material parameters, and outer boundary load can be kept consistent with the single-hole structure. The center distance between the holes is determined according to the actual bolt hole spacing or design parameters.
[0122] Based on the center positions of the two holes, the high-gradient neighborhoods of the hole edges on the outer sides of each hole are reconstructed, and the inter-hole ligament region between the two holes is taken as the key collocation region. Physical collocation points located in the ligament region are assigned high radial or combined distance weights, and incremental training is performed by combining the boundary constraints of the double-hole structure and a small amount of finite element hole edge reference stress.
[0123] After incremental training, the model can output the stress concentration distribution on the outer sides of the two holes and in the ligament region between the holes, and predict the circumferential stress curves at the edges of the left and right holes. Compared with training from random parameters, parameter transfer reuses the stress concentration patterns at the hole edges already learned in the single-hole structure, reducing the training burden on the multi-hole structure model.
[0124] Application Examples of Local Structures for Wind Power Bolt Connections
[0125] In wind turbine pitch bearings, yaw bearings, or flange bolted connection structures, a local region containing one or more adjacent bolt holes can be extracted from the full-ring multi-bolt layout, using local geometry, material parameters, bolt preload, or equivalent boundary loads as input parameters. For complex three-dimensional connection structures, local boundary displacements and traction can be extracted from a high-fidelity finite element model or a higher-level global model to form equivalent local boundary conditions.
[0126] The model trained using this invention can quickly output the stress distribution at the bolt hole edge and between adjacent holes, providing a basis for local mechanical response for hole edge fatigue risk analysis, preload state change assessment, structural optimization, and digital twin state update.
[0127] Simulation Device Examples
[0128] like Figure 8 As shown, the present invention also provides a simulation device for high gradient stress field of porous structures based on hole-guided physical constraint learning. The device includes a parameter acquisition module, a neighborhood construction module, a collocation generation module, a model building module, a physical constraint construction module, a boundary constraint construction module, a finite element weak supervision module, a model training module, and a stress field output module.
[0129] The parameter acquisition module is used to acquire geometric parameters, material parameters, boundary load parameters, and finite element hole edge reference data.
[0130] The neighborhood construction module is used to determine the computational domain, outer boundary, and hole boundary, and to construct the high gradient neighborhood of the hole edge based on the hole boundary.
[0131] The collocation generation module is used to generate high-density physical collocations in the high-gradient neighborhood of the hole edge and to generate background physical collocations in the background region.
[0132] The model building module is used to establish a hole-guided physical constraint learning model with spatial coordinates as input and displacement components as output, and to obtain strain and stress components based on automatic differentiation and material constitutive relations.
[0133] The physical constraint construction module is used to construct the equilibrium equation residuals, radial weighting functions, and radially weighted equilibrium equation losses.
[0134] The boundary constraint construction module is used to construct the outer boundary loss and the hole boundary loss, and update the weights of the two types of boundary losses based on the boundary residuals.
[0135] The finite element weak supervision module is used to extract the finite element reference stress at the hole edge and construct the weak supervision loss of the hole edge stress.
[0136] The model training module is used to train the model based on the total loss function and save the model parameters.
[0137] The stress field output module is used to output the displacement field, stress field, stress distribution at the hole edge, peak stress, and peak position.
[0138] Examples of computer equipment, storage media, and computer program products
[0139] like Figure 9As shown, the present invention also provides a computer device, including a processor 10, a memory 20, a communication interface 30, and a bus 40. The processor 10, the memory 20, and the communication interface 30 are connected via the bus 40. The memory 20 stores computer instructions, and when the processor 10 executes the computer instructions, it implements the above-described simulation method for high gradient stress fields of porous structures based on hole-guided physical constraint learning.
[0140] The processor 10 may be a central processing unit, a graphics processing unit, a network processor, or a combination thereof; the memory 20 may include random access memory, read-only memory, flash memory, disk storage, or other non-transient memory; the communication interface 30 is used to interact with a finite element calculation platform, a database, a monitoring system, or other computing devices.
[0141] The present invention also provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the above-described method.
[0142] The present invention also provides a computer program product, including a computer program or computer instructions; when the computer program or computer instructions are executed by a processor, the above-mentioned simulation method for high gradient stress field of porous structures based on hole-guided physical constraint learning is implemented.
[0143] According to the specific embodiments of the present invention described above, the present invention discloses a method, device, and storage medium for simulating high gradient stress fields of porous structures based on hole-guided physical constraint learning. The method includes: acquiring the geometric parameters, material parameters, boundary load parameters, and finite element hole edge reference data of the porous structure to be simulated; constructing a high gradient neighborhood of the hole edge based on the hole boundary, and arranging physical collocation points of different densities in the high gradient neighborhood of the hole edge and the background region respectively; constructing a hole-guided physical constraint learning model with spatial coordinates as input and displacement components as output, and obtaining strain components and stress components based on automatic differentiation and material constitutive relations; constructing radially weighted equilibrium equation constraints based on the distance from the physical collocation points to the hole boundary, constructing partition boundary constraints based on the outer boundary and the hole boundary, and introducing finite element hole edge reference data to construct weakly supervised hole edge stress constraints; training the model using the above constraints, and outputting the hole edge stress distribution, peak stress, peak position, and overall stress results; the method can also transfer the model parameters trained on a single-hole structure to a porous structure, and obtain the high gradient stress field of the porous structure through incremental training. This invention can enhance the physical constraint strength of the high gradient region at the hole edge, improve the predictive reliability of the peak stress and circumferential stress distribution at the hole edge, and reduce the computational cost of repeated finite element simulation under multiple working conditions.
[0144] The above description is merely a preferred embodiment of the present invention, and it should be understood that the present invention is not limited to the forms disclosed herein. Those skilled in the art can make equivalent substitutions or adjustments to the shape of the high-gradient neighborhood around the hole, the sampling method, the radial weight function, the boundary weight update method, the finite element reference stress type, the network structure, and the optimization algorithm without departing from the concept of the present invention; all such equivalent substitutions or adjustments should fall within the protection scope of the present invention.
Claims
1. A simulation method for high gradient stress field of porous structures based on hole-guided physical constraint learning, characterized in that: Includes the following steps: Step 1: Obtain the geometric parameters, material parameters, boundary load parameters, and finite element reference data for the hole edge of the structure to be simulated; Step 2: Determine the computational domain, outer boundary, and hole boundary of the hole structure based on the geometric parameters, and expand the hole boundary outward by a preset radial width to construct a high gradient neighborhood around the hole edge; Step 3: Arrange a first set of physical collocation points in the high gradient neighborhood of the hole edge, and arrange a second set of physical collocation points in the background region outside the high gradient neighborhood of the hole edge, and make the spatial density of the first set of physical collocation points greater than the spatial density of the second set of physical collocation points. Step 4: Establish a hole-guided physical constraint learning model with spatial coordinates as input and displacement components as output; calculate strain components based on automatic differentiation; and calculate stress components based on material constitutive relations. Step 5: Substitute the stress components into the static equilibrium equation to obtain the equilibrium equation residuals. Construct a radial weighting function based on the radial distance from the physical collocation point to the hole boundary, and weight the equilibrium equation residuals to obtain the radially weighted equilibrium equation loss. Step 6: Divide the boundary of the hole-containing structure into the outer boundary and the hole boundary, construct the outer boundary loss and the hole boundary loss respectively, and adaptively adjust the corresponding loss weights according to the relative magnitude of the outer boundary residual and the hole boundary residual during the training process to obtain the partition boundary loss. Step 7: Extract finite element reference stress data at several discrete angular positions at the hole boundary, and construct a weak supervision loss for hole edge stress based on the difference between the model-predicted stress and the finite element reference stress. Step 8: Construct a total loss function based on the radial weighted equilibrium equation loss, partition boundary loss, and weak supervision loss of hole edge stress, and use the total loss function to train the hole-guided physical constraint learning model; Step 9: Input the spatial coordinates of the area to be predicted into the trained hole-guided physical constraint learning model to obtain the displacement field, strain field and stress field of the hole-containing structure, and extract the stress distribution, peak stress and peak position at the hole edge. The geometric parameters include at least one of the following: hole radius, hole center position, number of holes, distance between holes, outer boundary dimension of structure, and structure thickness; the material parameters include elastic modulus and Poisson's ratio; the boundary load parameters include at least one of displacement boundary, far-field tensile load, equivalent surface force, and free boundary condition; the finite element hole edge reference data includes normal stress, shear stress, circumferential stress, or equivalent stress data at discrete angular positions of the hole edge; For a hole radius of A single-hole structure containing holes, wherein the high gradient neighborhood of the hole edge for: ; In equation (1), For point radial distance to the center of the hole The radial expansion width of the high gradient neighborhood around the hole edge; The location is determined based on the hole radius, structural dimensions, or the position of finite element stress gradient attenuation; for porous structures, the high gradient neighborhood of the hole edge also includes the inter-hole ligament region between adjacent holes. The strain component , and They are respectively Directional strain, Directional strain and Planar engineering shear strain, and satisfying: ; Under plane stress conditions, the stress components satisfy: ; ; ; In equations (4) to (6), For elastic modulus, Poisson's ratio; and They are respectively direction, Normal stress components in the direction, for Shear stress components in a plane; Ignoring physical strength Directional equilibrium equation residuals and Directional equilibrium equation residuals for: ; The radial weighting function decreases as the distance from the physical collocation point to the hole boundary increases, and satisfies: ; In equation (8), The hole edge constraint reinforcement coefficient is... These are the radial attenuation control parameters; Let be the radial distance from any physical point to the center of the hole. Where is the radius of the hole. Let be the radial distance from the physical point to the boundary of the circular hole. The radial weight corresponding to this physical collocation point; The loss of the radially weighted balance equation is: ; In equation (9), For the first Spatial coordinates of a physical point The radial distance from the physical point to the center of the hole. For the first Radial weights corresponding to each physical collocation point and The first At each physical coordinate point Directional equilibrium equation residuals and sum Directional equilibrium equation residuals This represents the total number of physical points.
2. The method according to claim 1, characterized in that: The partition boundary loss includes the outer boundary loss. and hole boundary loss The partition boundary loss satisfies: ; No. The outer boundary weights and hole boundary weights used in the training round are based on the first round. Mean boundary residuals after training rounds and Updated to: ; ; In equations (11) and (12), and The first The training round uses outer boundary loss weights and hole boundary loss weights; and The first The average value of the boundary residual amplitude at the outer boundary sampling point and the hole boundary sampling point after the round of training.
3. The method according to claim 1, characterized in that: The weak monitoring loss of the hole edge stress is: ; In equation (13), The number of reference points for the hole edge in the finite element method. Where is the radius of the hole. For the first The polar angle of a finite element reference point at the edge of a hole relative to the center of the hole; The type of the target stress component being monitored is indicated, wherein the target stress component is selected from normal stress, circumferential stress, shear stress, or equivalent stress; This indicates the position of the hole-guided physical constraint learning model at the hole edge. The predicted first Stress-like components, This represents the corresponding reference stress component obtained at the same hole edge location in the finite element model.
4. The method according to claim 1, characterized in that, The total loss function is: ; In equation (14), Let the total loss function be the physical constraint learning model guided by the hole. , and These are the weight coefficients for the radially weighted equilibrium equation loss, the partition boundary loss, and the weakly supervised hole edge stress loss, respectively. During training, a fully connected neural network with a sinusoidal activation function is used, and the Adam optimization algorithm is used to update the model parameters. In each training round, the physical collocation points and boundary collocation points are randomly resampled.
5. The method according to claim 1, characterized in that: It also includes training an initial hole-guided physical constraint learning model on a single-hole porous structure: using the parameters of the initial hole-guided physical constraint learning model as the initial parameters of the porous structure model; reconstructing the high gradient neighborhood of each hole edge according to the hole center position, hole radius and inter-hole distance of each hole in the porous structure, and rearranging high-density physical points in the hole edge region and the inter-hole ligament region; Incremental training of the model is performed based on the equilibrium equation residuals, boundary conditions, and finite element reference data of the porous structure.
6. The method according to claim 1, characterized in that: The high gradient stress field simulation results also include at least one of the following: circumferential stress curve at the hole edge, stress concentration region, stress concentration factor, peak angle position, equivalent stress, and full-field stress cloud map.
7. A simulation device for high gradient stress field of porous structures based on hole-guided physical constraint learning, characterized in that: include: Parameter acquisition module: used to acquire the geometric parameters, material parameters, boundary load parameters, and finite element hole edge reference data of the hole structure to be simulated; Neighborhood construction module: used to determine the computational domain, outer boundary, and hole boundary of the hole-containing structure, and to construct the high gradient neighborhood of the hole edge; Collocation generation module: used to generate high-density physical collocations in the high-gradient neighborhood of the hole edge, and to generate background physical collocations in the background region; Model building module: used to build a hole-guided physical constraint learning model with spatial coordinates as input and displacement components as output, and to obtain strain components and stress components; Physical constraint construction module: used to construct radially weighted equilibrium equation constraints; Boundary constraint construction module: used to construct the outer boundary loss and hole boundary loss, and adaptively adjust the weights of the outer boundary loss and hole boundary loss; Finite element weak supervision module: used to construct weak supervision constraints for hole edge stress based on finite element hole edge reference data; Model training module: used to construct the total loss function and train the hole-guided physical constraint learning model; Stress field output module: used to output the displacement field, stress field, stress distribution at the edge of the hole, peak stress and peak position of the hole structure.
8. A computer device for implementing the method as described in claims 1-7, characterized in that: The system includes a memory and a processor. The memory stores computer instructions, and when the processor executes the computer instructions, it implements the high gradient stress field simulation method for porous structures based on hole-guided physical constraint learning as described in any one of claims 1 to 7.
9. A computer-readable storage medium implementing the method as described in claims 1-7, characterized in that: The computer-readable storage medium stores computer instructions, which, when executed by a processor, implement the high gradient stress field simulation method for porous structures based on hole-guided physical constraint learning as described in any one of claims 1 to 7.
10. A computer program product implementing the method as described in claims 1-7, characterized in that: It includes a computer program or computer instructions, which, when executed by a processor, implement the high gradient stress field simulation method for porous structures based on hole-guided physical constraint learning as described in any one of claims 1 to 7.
